The infrastructure that isn't
If you've heard of Cursor or Claude Code, you might picture a complex technical setup behind the scenes — API integrations, custom models, layers of configuration. The reality is almost embarrassingly simple.
Cursor stores your coding rules in a file called .cursorrules. Claude Code reads a file called CLAUDE.md. GitHub Copilot uses .github/copilot-instructions.md. Open any of these files and you'll find plain English text — instructions written exactly as you'd write them in a chat window. No code. No schema. No special syntax.
server.py config.json requirements.txt deploy.sh api_client.py vector_store/ embeddings.pkl index.faiss models/ fine_tuned_v3/
agent.md
These files are a system prompt saved to disk. That's the entire secret. Every AI tool that feels "smart" or "configured" is reading a text file before it talks to you. And you can write that file yourself.
What is a system prompt?
Every AI conversation has a hidden opening message that shapes everything that follows. Developers call it a system prompt. It runs before you type a single word, and it tells the AI: who you are, how to behave, what you know, what to avoid.
When you open a chat with a customer service bot that only answers questions about shoes, that constraint is written in a system prompt. When Claude formats its responses a certain way, or stays on topic, or refuses to discuss certain things — that's a system prompt at work. You just don't see it.
Writing one is the same as writing a memo. You describe a role. You list some rules. You give context. Here is a complete, working agent:
You are a writing assistant for a small consulting firm. TONE - Professional but not stiff - Use plain English, no jargon - Write for a general business audience RULES - Never use bullet points for single items - Keep paragraphs to 3–4 sentences maximum - When in doubt, make it shorter WHAT TO AVOID - Filler phrases like "It's worth noting that" or "As mentioned above" - Passive voice unless it genuinely fits - Starting a sentence with "However" more than once per document
That's it. Copy this into the system prompt field of any AI tool that exposes one — Claude, ChatGPT, Copilot, Gemini — and every response you get will follow these rules. You haven't written code. You haven't touched an API. You've built an agent.
The key insight: Cursor's .cursorrules and Claude's CLAUDE.md look exactly like the long instruction you'd paste into a chat window. The only difference is that they're saved to a file — so the AI reads them automatically, every time, without you having to paste anything.
Giving your agent a memory
A system prompt handles identity and rules. But what about knowledge? What if your agent needs to know your organization's terminology, the structure of a report you're working on, or the particular way your team writes to clients?
The answer is reference files: separate markdown documents that your agent is instructed to consult before acting. You describe these files inside the system prompt itself, using plain language:
BEFORE WRITING ANYTHING Read and apply the rules in these files: - style-guide.md — tone, vocabulary, formatting - product-descriptions.md — approved product names and descriptions - contact-list.md — team names and titles Never invent product names or team titles that aren't in these files.
This is the mechanism behind every impressive-looking AI assistant. The AI doesn't know your company from training — it reads a file you wrote. The "memory" is just a markdown document in the same folder.
You can point to as many files as you need. A style guide. A glossary. A list of FAQs. A summary of the project background. Each file extends what the agent knows without changing the core instruction.
One practical limit: AI tools have a context window — a maximum amount of text they can hold in mind at once. Very large reference files may get truncated. Keep individual files focused and reasonably sized. A 500-word style guide works better than a 20,000-word style manual.
Folders tell the story
Once you have more than one reference file, how you organize them starts to matter. Folder structure is a way of describing scope — which rules apply everywhere, and which apply only in a specific context.
The convention used by professional tools is simple: files at the root of a project apply to everything. Files inside a subfolder apply only when working in that area. You can describe this structure in plain text inside your system prompt:
PROJECT LAYOUT /style-guide.md — applies to all writing in this project /tone-of-voice.md — applies to all writing in this project /proposals/brief.md — applies only when working on proposals /proposals/template.md — the standard proposal structure /emails/signature.md — applies only when drafting emails /emails/common-replies.md — approved language for common situations
The AI doesn't automatically see the folder structure — you describe it. But once described, the agent understands which context it's in and applies the right rules. This is exactly how Claude Code uses CLAUDE.md: a root file for the whole project, and subfolder files for specific areas.
Your first text agent
Let's make this concrete. Say you want an agent that helps you write consistently — proposals, client emails, and status reports — all following the same voice, with no re-briefing required each session.
Here is everything you need:
What you need
agent.md— role, rules, what to readstyle-guide.md— tone, vocabulary, formattingreferences/— a folder with supporting facts- Any AI tool that accepts a system prompt
What you don't need
- A server or hosting
- An API key or developer account
- Any programming language
- More than an hour to set up
Paste the contents of agent.md into the system prompt field of your AI tool. Add the contents of style-guide.md and any reference files beneath it, separated by a heading so the agent knows which is which. Start a conversation. The agent will follow the rules in those files for the rest of the session — and whenever you start a new session, paste the same files in again.
Once you've done this a few times, you'll notice the natural next step: saving these files somewhere you can reach them quickly, so pasting them in becomes a single copy-paste instead of reassembling them each time. That's precisely what Cursor and Claude Code automated — nothing more.
What makes this an agent
The word "agent" gets used loosely. In this context it means something specific: an AI that behaves consistently because its instructions are stable — not because you remembered to re-explain the rules each time.
When you paste a long instruction into a chat window once, you've created a one-session agent. Useful, but fragile — the rules disappear when the conversation ends. When you save those instructions to files and point the AI at them every session, you've built something that persists. Something you can refine over time. Something you can hand to a colleague.
This is what professionals are actually doing when they configure tools like Cursor or Claude Code. The barrier wasn't technical — it was that nobody explained it wasn't technical. You just need a text file and the habit of using it.
Next step: Want to take this to a coding context — where the agent understands your project structure, your coding conventions, and your preferred libraries? That's Part 2 of this guide. The principle is identical. The files just contain different instructions.